The Earnings Print: Sunday Roundup

The Earnings Print - We’ll be publishing this actively around major earnings windows.

There is too much happening across the AI supply chain for any one investor to digest call by call. So we’re doing the cross-read. The biggest signals across compute, memory, networking, data centers, and power, synthesized into one easy read for you.

Around twenty important suppliers reported earnings last week. The market has started to move with them.

Neoclouds had a monster week. Our GPU Cloud/Neocloud basket, CRWV, NBIS, IREN, APLD, WULF, and CORZ, is up almost 30% over the past 7 days.

We’ve noticed that the companies underneath the hyperscalers are now telling the story before the hyperscalers say it explicitly: power is becoming the binding constraint, and photonics is becoming the next memory-like bottleneck.

Also: AI infrastructure is starting to underwrite like an institutional asset class, with contracted load, contracted capacity, investment-grade debt, and multi-year backlog replacing speculative capacity stories.

This is the week’s read: the bottleneck moved, and the market is still catching up.

The strongest signal this week came from power.

Vistra signed long-term PPAs with Meta for approximately 2,600 MW of nuclear capacity at its PJM sites, turning hyperscaler power demand into contracted revenue.

Hut 8 signed a 352 MW IT lease, 500 MW utility, on a 15-year triple-net structure with an AA-or-higher counterparty. The base contract value is $9.8B before escalators and $25B inclusive. It also issued what it described as the first investment-grade construction bond for a single-sponsored data center project.

Core Scientific said a hyperscaler exclusivity period at Pecos and Muskogee expired, and three hyperscalers engaged on those same sites within days. It also raised the gross margin target on its CoreWeave contract from 75 to 80% to 80 to 85%.

Exelon now expects transmission rate base growth of 16% through 2029 and warned publicly about 2028 reliability risk in PJM. NRG sized the ERCOT large-load pipeline at 36 GW by 2033 and disclosed 2 GW of in-fleet uprate opportunities.

Here’s the pattern: AI demand is moving from GPU orders into power contracts, transmission planning, grid reliability, nuclear procurement, and behind-the-meter capacity.

You can’t think of power as a second-order constraint anymore. It is becoming the thing that determines where compute can exist.

Two financing firsts happened in the same week.

CoreWeave closed an A-minus-equivalent delayed-draw term loan against HPC infrastructure, the first investment-grade rating ever achieved on this asset class.

Hut 8 issued the first investment-grade construction bond for a single-sponsored data center.

Both companies also increased multi-year contract visibility. CoreWeave said more than 75% of its $30B 2027 ARR target is already contracted, excluding renewals. Total contracted backlog now sits near $100B.

This changes the capital base of the AI buildout.

The first phase was capacity-starved and speculative. The next phase is becoming more institutional: long-duration contracts, contracted power, investment-grade debt, and infrastructure-style underwriting.

When construction risk on a data center can rate investment-grade on day one (!!), the asset class has taken on a life of its own.

The second major constraint is optical networking. From earnings this week:

Lumentum’s EML supply-demand shortfall widened to 25 to 30%, worse than the 20% shortfall it disclosed a quarter ago, despite adding roughly 40% more unit capacity.

Applied Optoelectronics quoted 21-to-24-month equipment lead times for indium phosphide laser production. It also disclosed actual demand of $1.4 to $1.5B against a $1.2B revenue target.

MACOM raised its FY26 data center growth target from 35 to 40% to more than 60%, posted a 1.5x book-to-bill, and took an 11% equity stake in IQE for £45M to lock down epitaxial supply.

Veeco disclosed more than $250M in indium phosphide laser orders and a 10x expansion target for SPECTOR IBD capacity by early 2027.

Four independent companies all hinting at the same shortage.

The market understands HBM scarcity. It is only beginning to understand optical scarcity.

The oddest part of the week was what companies did not say.

Across the earnings calls we read, no Tier-1 company quantified specific ASP increases despite broadly describing tight supply.

Lumentum and Applied Optoelectronics both acknowledged widening shortfalls without disclosing pricing direction.

AMD was asked directly about taking price into rising memory costs and did not quantify an increase.

There are two possible explanations.

We would put more weight on the second.

Tight supply is interesting, but pricing power is investable. The winners are the companies that can convert scarcity into better margins.

Physical AI is starting to show up in reported numbers.

Symbotic disclosed 1 million autonomous bot-miles per day. Ouster shipped native-color lidar silicon and moved into industrial safety. MP Materials framed physical AI as a rare-earth demand inflection, while noting that Japan has locked up nearly all of Lynas’ NdPr output.

This is earlier than the power and photonics story, but robotics is no longer just a demo narrative.

Not everything tightened cleanly.

Eaton missed on Electrical Americas margins, citing higher input costs and the cost of delivering higher volumes.

That is worth watching.

The demand picture at the electrical and thermal layer still looks strong. Boyd Thermal’s liquid-cooling backlog doubled in six months. But Eaton’s margin miss is a reminder that volume growth is not the same thing as clean operating leverage.

The next leg of the AI cycle is shifting toward owners of scarce power, contracted data center capacity, optical bottlenecks and upstream materials.

But demand alone is not enough.

It is whether demand converts into one of three things:

The compute supply chain is locking down further out, faster. And the binding constraints have moved to power and photonics.

Up next: NVDA in 10 days

Tessara's earnings intelligence ships pre-call synthesis, post-call synthesis, and our variant calls (where our model disagrees with consensus) for every major AI infrastructure print. NVDA drops in 10 days. We're building the pre-call read now. Apply for early access →

Cheers,

Teng